/*
+ * Copyright (c) 2000-2012 Apple Inc. All rights reserved.
*
* @APPLE_OSREFERENCE_LICENSE_HEADER_START@
*
#include <mach_rt.h>
#include <mach_kdp.h>
+#include <kdp/kdp_internal.h>
#include <mach_ldebug.h>
#include <gprof.h>
#include <kern/machine.h>
#include <kern/pms.h>
#include <kern/misc_protos.h>
-#include <kern/etimer.h>
+#include <kern/timer_call.h>
#include <kern/kalloc.h>
#include <kern/queue.h>
+#include <prng/random.h>
#include <vm/vm_map.h>
#include <vm/vm_kern.h>
#include <profiling/profile-mk.h>
+#include <i386/bit_routines.h>
#include <i386/proc_reg.h>
#include <i386/cpu_threads.h>
#include <i386/mp_desc.h>
#include <sys/kdebug.h>
+#include <console/serial_protos.h>
+
#if MP_DEBUG
#define PAUSE delay(1000000)
#define DBG(x...) kprintf(x)
void slave_boot_init(void);
void i386_cpu_IPI(int cpu);
+#if MACH_KDP
static void mp_kdp_wait(boolean_t flush, boolean_t isNMI);
-static void mp_rendezvous_action(void);
-static void mp_broadcast_action(void);
+#endif /* MACH_KDP */
+static boolean_t mp_safe_spin_lock(usimple_lock_t lock);
+#if MACH_KDP
static boolean_t cpu_signal_pending(int cpu, mp_event_t event);
+#endif /* MACH_KDP */
static int NMIInterruptHandler(x86_saved_state_t *regs);
boolean_t smp_initialized = FALSE;
volatile boolean_t pmap_tlb_flush_timeout = FALSE;
decl_simple_lock_data(,mp_kdp_lock);
+decl_simple_lock_data(,debugger_callback_lock);
+struct debugger_callback *debugger_callback = NULL;
+
decl_lck_mtx_data(static, mp_cpu_boot_lock);
lck_mtx_ext_t mp_cpu_boot_lock_ext;
decl_lck_mtx_data(static, mp_bc_lock);
lck_mtx_ext_t mp_bc_lock_ext;
static volatile int debugger_cpu = -1;
-volatile long NMIPI_acks = 0;
+volatile long NMIPI_acks = 0;
+volatile long NMI_count = 0;
+
+extern void NMI_cpus(void);
static void mp_cpus_call_init(void);
-static void mp_cpus_call_cpu_init(void);
static void mp_cpus_call_action(void);
static void mp_call_PM(void);
+static boolean_t mp_cpus_call_wait_timeout = FALSE;
+
char mp_slave_stack[PAGE_SIZE] __attribute__((aligned(PAGE_SIZE))); // Temp stack for slave init
/* PAL-related routines */
{
simple_lock_init(&mp_kdp_lock, 0);
simple_lock_init(&mp_rv_lock, 0);
+ simple_lock_init(&debugger_callback_lock, 0);
lck_grp_attr_setdefault(&smp_lck_grp_attr);
lck_grp_init(&smp_lck_grp, "i386_smp", &smp_lck_grp_attr);
lck_mtx_init_ext(&mp_cpu_boot_lock, &mp_cpu_boot_lock_ext, &smp_lck_grp, LCK_ATTR_NULL);
DBGLOG_CPU_INIT(master_cpu);
mp_cpus_call_init();
- mp_cpus_call_cpu_init();
+ mp_cpus_call_cpu_init(master_cpu);
if (PE_parse_boot_argn("TSC_sync_margin",
&TSC_sync_margin, sizeof(TSC_sync_margin))) {
* Initialize (or re-initialize) the descriptor tables for this cpu.
* Propagate processor mode to slave.
*/
- if (cpu_mode_is64bit())
- cpu_desc_init64(cpu_datap(slot_num));
- else
- cpu_desc_init(cpu_datap(slot_num));
+ cpu_desc_init64(cpu_datap(slot_num));
/* Serialize use of the slave boot stack, etc. */
lck_mtx_lock(&mp_cpu_boot_lock);
#endif /* MP_DEBUG */
+/*
+ * Note: called with NULL state when polling for TLB flush and cross-calls.
+ */
int
cpu_signal_handler(x86_saved_state_t *regs)
{
+#if !MACH_KDP
+#pragma unused (regs)
+#endif /* !MACH_KDP */
int my_cpu;
volatile int *my_word;
DBGLOG(cpu_handle,my_cpu,MP_TLB_FLUSH);
i_bit_clear(MP_TLB_FLUSH, my_word);
pmap_update_interrupt();
- } else if (i_bit(MP_AST, my_word)) {
- DBGLOG(cpu_handle,my_cpu,MP_AST);
- i_bit_clear(MP_AST, my_word);
- ast_check(cpu_to_processor(my_cpu));
- } else if (i_bit(MP_RENDEZVOUS, my_word)) {
- DBGLOG(cpu_handle,my_cpu,MP_RENDEZVOUS);
- i_bit_clear(MP_RENDEZVOUS, my_word);
- mp_rendezvous_action();
- } else if (i_bit(MP_BROADCAST, my_word)) {
- DBGLOG(cpu_handle,my_cpu,MP_BROADCAST);
- i_bit_clear(MP_BROADCAST, my_word);
- mp_broadcast_action();
} else if (i_bit(MP_CHUD, my_word)) {
DBGLOG(cpu_handle,my_cpu,MP_CHUD);
i_bit_clear(MP_CHUD, my_word);
i_bit_clear(MP_CALL_PM, my_word);
mp_call_PM();
}
+ if (regs == NULL) {
+ /* Called to poll only for cross-calls and TLB flush */
+ break;
+ } else if (i_bit(MP_AST, my_word)) {
+ DBGLOG(cpu_handle,my_cpu,MP_AST);
+ i_bit_clear(MP_AST, my_word);
+ ast_check(cpu_to_processor(my_cpu));
+ }
} while (*my_word);
return 0;
}
+extern void kprintf_break_lock(void);
static int
NMIInterruptHandler(x86_saved_state_t *regs)
{
- void *stackptr;
+ void *stackptr;
if (panic_active() && !panicDebugging) {
if (pmsafe_debug)
}
atomic_incl(&NMIPI_acks, 1);
+ atomic_incl(&NMI_count, 1);
sync_iss_to_iks_unconditionally(regs);
-#if defined (__i386__)
- __asm__ volatile("movl %%ebp, %0" : "=m" (stackptr));
-#elif defined (__x86_64__)
__asm__ volatile("movq %%rbp, %0" : "=m" (stackptr));
-#endif
if (cpu_number() == debugger_cpu)
- goto NMExit;
+ goto NMExit;
if (spinlock_timed_out) {
char pstr[192];
snprintf(&pstr[0], sizeof(pstr), "Panic(CPU %d): NMIPI for spinlock acquisition timeout, spinlock: %p, spinlock owner: %p, current_thread: %p, spinlock_owner_cpu: 0x%x\n", cpu_number(), spinlock_timed_out, (void *) spinlock_timed_out->interlock.lock_data, current_thread(), spinlock_owner_cpu);
panic_i386_backtrace(stackptr, 64, &pstr[0], TRUE, regs);
+ } else if (mp_cpus_call_wait_timeout) {
+ char pstr[192];
+ snprintf(&pstr[0], sizeof(pstr), "Panic(CPU %d): Unresponsive processor, this CPU timed-out during cross-call\n", cpu_number());
+ panic_i386_backtrace(stackptr, 64, &pstr[0], TRUE, regs);
} else if (pmap_tlb_flush_timeout == TRUE) {
char pstr[128];
snprintf(&pstr[0], sizeof(pstr), "Panic(CPU %d): Unresponsive processor (this CPU did not acknowledge interrupts) TLB state:0x%x\n", cpu_number(), current_cpu_datap()->cpu_tlb_invalid);
panic_i386_backtrace(stackptr, 48, &pstr[0], TRUE, regs);
- }
+ }
#if MACH_KDP
if (pmsafe_debug && !kdp_snapshot)
pmSafeMode(¤t_cpu_datap()->lcpu, PM_SAFE_FL_SAFE);
current_cpu_datap()->cpu_NMI_acknowledged = TRUE;
- mp_kdp_wait(FALSE, pmap_tlb_flush_timeout || spinlock_timed_out || panic_active());
+ i_bit_clear(MP_KDP, ¤t_cpu_datap()->cpu_signals);
+ if (pmap_tlb_flush_timeout ||
+ spinlock_timed_out ||
+ mp_cpus_call_wait_timeout ||
+ panic_active()) {
+ mp_kdp_wait(FALSE, TRUE);
+ } else if (!mp_kdp_trap &&
+ !mp_kdp_is_NMI &&
+ virtualized && (debug_boot_arg & DB_NMI)) {
+ /*
+ * Under a VMM with the debug boot-arg set, drop into kdp.
+ * Since an NMI is involved, there's a risk of contending with
+ * a panic. And side-effects of NMIs may result in entry into,
+ * and continuing from, the debugger being unreliable.
+ */
+ if (__sync_bool_compare_and_swap(&mp_kdp_is_NMI, FALSE, TRUE)) {
+ kprintf_break_lock();
+ kprintf("Debugger entry requested by NMI\n");
+ kdp_i386_trap(T_DEBUG, saved_state64(regs), 0, 0);
+ printf("Debugger entry requested by NMI\n");
+ mp_kdp_is_NMI = FALSE;
+ } else {
+ mp_kdp_wait(FALSE, FALSE);
+ }
+ } else {
+ mp_kdp_wait(FALSE, FALSE);
+ }
if (pmsafe_debug && !kdp_snapshot)
pmSafeMode(¤t_cpu_datap()->lcpu, PM_SAFE_FL_NORMAL);
#endif
}
}
+void
+NMI_cpus(void)
+{
+ unsigned int cpu;
+ boolean_t intrs_enabled;
+ uint64_t tsc_timeout;
+
+ intrs_enabled = ml_set_interrupts_enabled(FALSE);
+
+ for (cpu = 0; cpu < real_ncpus; cpu++) {
+ if (!cpu_datap(cpu)->cpu_running)
+ continue;
+ cpu_datap(cpu)->cpu_NMI_acknowledged = FALSE;
+ cpu_NMI_interrupt(cpu);
+ tsc_timeout = !machine_timeout_suspended() ?
+ rdtsc64() + (1000 * 1000 * 1000 * 10ULL) :
+ ~0ULL;
+ while (!cpu_datap(cpu)->cpu_NMI_acknowledged) {
+ handle_pending_TLB_flushes();
+ cpu_pause();
+ if (rdtsc64() > tsc_timeout)
+ panic("NMI_cpus() timeout cpu %d", cpu);
+ }
+ cpu_datap(cpu)->cpu_NMI_acknowledged = FALSE;
+ }
+
+ ml_set_interrupts_enabled(intrs_enabled);
+}
+
static void (* volatile mp_PM_func)(void) = NULL;
static void
i386_cpu_IPI(cpu);
if (mode == SYNC) {
again:
- tsc_timeout = rdtsc64() + (1000*1000*1000);
+ tsc_timeout = !machine_timeout_suspended() ?
+ rdtsc64() + (1000*1000*1000) :
+ ~0ULL;
while (i_bit(event, signals) && rdtsc64() < tsc_timeout) {
cpu_pause();
}
}
/*
- * Send event to all running cpus.
- * Called with the topology locked.
+ * Helper function called when busy-waiting: panic if too long
+ * a TSC-based time has elapsed since the start of the spin.
*/
-void
-i386_signal_cpus(mp_event_t event, mp_sync_t mode)
+static boolean_t
+mp_spin_timeout(uint64_t tsc_start)
{
- unsigned int cpu;
- unsigned int my_cpu = cpu_number();
+ uint64_t tsc_timeout;
- assert(hw_lock_held((hw_lock_t)&x86_topo_lock));
+ cpu_pause();
+ if (machine_timeout_suspended())
+ return FALSE;
- for (cpu = 0; cpu < real_ncpus; cpu++) {
- if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
- continue;
- i386_signal_cpu(cpu, event, mode);
- }
+ /*
+ * The timeout is 4 * the spinlock timeout period
+ * unless we have serial console printing (kprintf) enabled
+ * in which case we allow an even greater margin.
+ */
+ tsc_timeout = disable_serial_output ? LockTimeOutTSC << 2
+ : LockTimeOutTSC << 4;
+ return (rdtsc64() > tsc_start + tsc_timeout);
}
/*
- * Return the number of running cpus.
- * Called with the topology locked.
+ * Helper function to take a spinlock while ensuring that incoming IPIs
+ * are still serviced if interrupts are masked while we spin.
+ * Returns current interrupt state.
*/
-int
-i386_active_cpus(void)
+static boolean_t
+mp_safe_spin_lock(usimple_lock_t lock)
{
- unsigned int cpu;
- unsigned int ncpus = 0;
-
- assert(hw_lock_held((hw_lock_t)&x86_topo_lock));
-
- for (cpu = 0; cpu < real_ncpus; cpu++) {
- if (cpu_datap(cpu)->cpu_running)
- ncpus++;
- }
- return(ncpus);
+ if (ml_get_interrupts_enabled()) {
+ simple_lock(lock);
+ return TRUE;
+ } else {
+ uint64_t tsc_spin_start = rdtsc64();
+ while (!simple_lock_try(lock)) {
+ cpu_signal_handler(NULL);
+ if (mp_spin_timeout(tsc_spin_start)) {
+ uint32_t lock_cpu;
+ uintptr_t lowner = (uintptr_t)
+ lock->interlock.lock_data;
+ spinlock_timed_out = lock;
+ lock_cpu = spinlock_timeout_NMI(lowner);
+ panic("mp_safe_spin_lock() timed out,"
+ " lock: %p, owner thread: 0x%lx,"
+ " current_thread: %p, owner on CPU 0x%x",
+ lock, lowner,
+ current_thread(), lock_cpu);
+ }
+ }
+ return FALSE;
+ }
}
/*
*/
static void
-mp_rendezvous_action(void)
+mp_rendezvous_action(__unused void *null)
{
- boolean_t intrs_enabled;
+ boolean_t intrs_enabled;
+ uint64_t tsc_spin_start;
/* setup function */
if (mp_rv_setup_func != NULL)
/* spin on entry rendezvous */
atomic_incl(&mp_rv_entry, 1);
+ tsc_spin_start = rdtsc64();
+
while (mp_rv_entry < mp_rv_ncpus) {
/* poll for pesky tlb flushes if interrupts disabled */
if (!intrs_enabled)
handle_pending_TLB_flushes();
- cpu_pause();
+ if (mp_spin_timeout(tsc_spin_start)) {
+ panic("mp_rv_action() entry: %ld of %d responses, start: 0x%llx, cur: 0x%llx", mp_rv_entry, mp_rv_ncpus, tsc_spin_start, rdtsc64());
+ }
}
/* action function */
/* spin on exit rendezvous */
atomic_incl(&mp_rv_exit, 1);
+ tsc_spin_start = rdtsc64();
while (mp_rv_exit < mp_rv_ncpus) {
if (!intrs_enabled)
handle_pending_TLB_flushes();
- cpu_pause();
+ if (mp_spin_timeout(tsc_spin_start))
+ panic("mp_rv_action() exit: %ld of %d responses, start: 0x%llx, cur: 0x%llx", mp_rv_exit, mp_rv_ncpus, tsc_spin_start, rdtsc64());
}
/* teardown function */
void (*teardown_func)(void *),
void *arg)
{
+ uint64_t tsc_spin_start;
if (!smp_initialized) {
if (setup_func != NULL)
}
/* obtain rendezvous lock */
- simple_lock(&mp_rv_lock);
+ (void) mp_safe_spin_lock(&mp_rv_lock);
/* set static function pointers */
mp_rv_setup_func = setup_func;
* signal other processors, which will call mp_rendezvous_action()
* with interrupts disabled
*/
- simple_lock(&x86_topo_lock);
- mp_rv_ncpus = i386_active_cpus();
- i386_signal_cpus(MP_RENDEZVOUS, ASYNC);
- simple_unlock(&x86_topo_lock);
+ mp_rv_ncpus = mp_cpus_call(CPUMASK_OTHERS, NOSYNC, &mp_rendezvous_action, NULL) + 1;
/* call executor function on this cpu */
- mp_rendezvous_action();
+ mp_rendezvous_action(NULL);
/*
* Spin for everyone to complete.
* This is necessary to ensure that all processors have proceeded
* from the exit barrier before we release the rendezvous structure.
*/
+ tsc_spin_start = rdtsc64();
while (mp_rv_complete < mp_rv_ncpus) {
- cpu_pause();
+ if (mp_spin_timeout(tsc_spin_start))
+ panic("mp_rendezvous() timeout: %ld of %d responses, start: 0x%llx, cur: 0x%llx", mp_rv_complete, mp_rv_ncpus, tsc_spin_start, rdtsc64());
}
/* Tidy up */
void (*func)(void *,void *); /* routine to call */
void *arg0; /* routine's 1st arg */
void *arg1; /* routine's 2nd arg */
- volatile long *countp; /* completion counter */
+ cpumask_t *maskp; /* completion response mask */
} mp_call_t;
return intrs_enabled;
}
+void
+mp_cpus_NMIPI(cpumask_t cpu_mask) {
+ unsigned int cpu, cpu_bit;
+ uint64_t deadline;
+
+ for (cpu = 0, cpu_bit = 1; cpu < real_ncpus; cpu++, cpu_bit <<= 1) {
+ if (cpu_mask & cpu_bit)
+ cpu_NMI_interrupt(cpu);
+ }
+ deadline = mach_absolute_time() + (LockTimeOut);
+ while (mach_absolute_time() < deadline)
+ cpu_pause();
+}
+
+#if MACH_ASSERT
static inline boolean_t
mp_call_head_is_locked(mp_call_queue_t *cqp)
{
return !ml_get_interrupts_enabled() &&
hw_lock_held((hw_lock_t)&cqp->lock);
}
+#endif
static inline void
mp_call_head_unlock(mp_call_queue_t *cqp, boolean_t intrs_enabled)
}
/*
- * Called by each processor to add call buffers to the free list
+ * Called at processor registration to add call buffers to the free list
* and to initialize the per-cpu call queue.
- * Also called but ignored on slave processors on re-start/wake.
*/
-static void
-mp_cpus_call_cpu_init(void)
+void
+mp_cpus_call_cpu_init(int cpu)
{
int i;
- mp_call_queue_t *cqp = &mp_cpus_call_head[cpu_number()];
+ mp_call_queue_t *cqp = &mp_cpus_call_head[cpu];
mp_call_t *callp;
- if (cqp->queue.next != NULL)
- return; /* restart/wake case: called already */
-
simple_lock_init(&cqp->lock, 0);
queue_init(&cqp->queue);
for (i = 0; i < MP_CPUS_CALL_BUFS_PER_CPU; i++) {
mp_call_free(callp);
}
- DBG("mp_cpus_call_init() done on cpu %d\n", cpu_number());
+ DBG("mp_cpus_call_init(%d) done\n", cpu);
}
/*
mp_call_head_unlock(cqp, intrs_enabled);
KERNEL_DEBUG_CONSTANT(
TRACE_MP_CPUS_CALL_ACTION,
- call.func, call.arg0, call.arg1, call.countp, 0);
+ VM_KERNEL_UNSLIDE(call.func), VM_KERNEL_UNSLIDE_OR_PERM(call.arg0),
+ VM_KERNEL_UNSLIDE_OR_PERM(call.arg1), VM_KERNEL_ADDRPERM(call.maskp), 0);
call.func(call.arg0, call.arg1);
(void) mp_call_head_lock(cqp);
}
- if (call.countp != NULL)
- atomic_incl(call.countp, 1);
+ if (call.maskp != NULL)
+ i_bit_set(cpu_number(), call.maskp);
}
mp_call_head_unlock(cqp, intrs_enabled);
}
(void (*)(void *,void *))action_func,
arg,
NULL,
- NULL,
NULL);
}
static void
mp_cpus_call_wait(boolean_t intrs_enabled,
- long mp_cpus_signals,
- volatile long *mp_cpus_calls)
+ cpumask_t cpus_called,
+ cpumask_t *cpus_responded)
{
mp_call_queue_t *cqp;
+ uint64_t tsc_spin_start;
+ assert(ml_get_interrupts_enabled() == 0 || get_preemption_level() != 0);
cqp = &mp_cpus_call_head[cpu_number()];
- while (*mp_cpus_calls < mp_cpus_signals) {
+ tsc_spin_start = rdtsc64();
+ while (*cpus_responded != cpus_called) {
if (!intrs_enabled) {
/* Sniffing w/o locking */
if (!queue_empty(&cqp->queue))
mp_cpus_call_action();
- handle_pending_TLB_flushes();
+ cpu_signal_handler(NULL);
+ }
+ if (mp_spin_timeout(tsc_spin_start)) {
+ cpumask_t cpus_unresponsive;
+
+ mp_cpus_call_wait_timeout = TRUE;
+ cpus_unresponsive = cpus_called & ~(*cpus_responded);
+ mp_cpus_NMIPI(cpus_unresponsive);
+ panic("mp_cpus_call_wait() timeout, cpus: 0x%llx",
+ cpus_unresponsive);
}
- cpu_pause();
}
}
void (*action_func)(void *, void *),
void *arg0,
void *arg1,
- cpumask_t *cpus_calledp,
- cpumask_t *cpus_notcalledp)
+ cpumask_t *cpus_calledp)
{
- cpu_t cpu;
+ cpu_t cpu = 0;
boolean_t intrs_enabled = FALSE;
boolean_t call_self = FALSE;
cpumask_t cpus_called = 0;
- cpumask_t cpus_notcalled = 0;
- long mp_cpus_signals = 0;
- volatile long mp_cpus_calls = 0;
+ cpumask_t cpus_responded = 0;
+ long cpus_call_count = 0;
+ uint64_t tsc_spin_start;
+ boolean_t topo_lock;
KERNEL_DEBUG_CONSTANT(
TRACE_MP_CPUS_CALL | DBG_FUNC_START,
- cpus, mode, VM_KERNEL_UNSLIDE(action_func), arg0, arg1);
+ cpus, mode, VM_KERNEL_UNSLIDE(action_func), VM_KERNEL_UNSLIDE_OR_PERM(arg0), VM_KERNEL_UNSLIDE_OR_PERM(arg1));
if (!smp_initialized) {
if ((cpus & CPUMASK_SELF) == 0)
/*
* Queue the call for each non-local requested cpu.
- * The topo lock is not taken. Instead we sniff the cpu_running state
- * and then re-check it after taking the call lock. A cpu being taken
- * offline runs the action function after clearing the cpu_running.
+ * This is performed under the topo lock to prevent changes to
+ * cpus online state and to prevent concurrent rendezvouses --
+ * although an exception is made if we're calling only the master
+ * processor since that always remains active. Note: this exception
+ * is expected for longterm timer nosync cross-calls to the master cpu.
*/
+ mp_disable_preemption();
+ intrs_enabled = ml_get_interrupts_enabled();
+ topo_lock = (cpus != cpu_to_cpumask(master_cpu));
+ if (topo_lock) {
+ ml_set_interrupts_enabled(FALSE);
+ (void) mp_safe_spin_lock(&x86_topo_lock);
+ }
for (cpu = 0; cpu < (cpu_t) real_ncpus; cpu++) {
if (((cpu_to_cpumask(cpu) & cpus) == 0) ||
!cpu_datap(cpu)->cpu_running)
continue;
+ tsc_spin_start = rdtsc64();
if (cpu == (cpu_t) cpu_number()) {
/*
* We don't IPI ourself and if calling asynchronously,
* we defer our call until we have signalled all others.
*/
call_self = TRUE;
- cpus_called |= cpu_to_cpumask(cpu);
if (mode == SYNC && action_func != NULL) {
KERNEL_DEBUG_CONSTANT(
TRACE_MP_CPUS_CALL_LOCAL,
VM_KERNEL_UNSLIDE(action_func),
- arg0, arg1, 0, 0);
+ VM_KERNEL_UNSLIDE_OR_PERM(arg0), VM_KERNEL_UNSLIDE_OR_PERM(arg1), 0, 0);
action_func(arg0, arg1);
}
} else {
/*
* Here to queue a call to cpu and IPI.
- * Spinning for request buffer unless NOSYNC.
*/
mp_call_t *callp = NULL;
mp_call_queue_t *cqp = &mp_cpus_call_head[cpu];
+ boolean_t intrs_inner;
queue_call:
if (callp == NULL)
callp = mp_call_alloc();
- intrs_enabled = mp_call_head_lock(cqp);
- if (!cpu_datap(cpu)->cpu_running) {
- mp_call_head_unlock(cqp, intrs_enabled);
- continue;
- }
- if (mode == NOSYNC) {
- if (callp == NULL) {
- cpus_notcalled |= cpu_to_cpumask(cpu);
- mp_call_head_unlock(cqp, intrs_enabled);
- KERNEL_DEBUG_CONSTANT(
- TRACE_MP_CPUS_CALL_NOBUF,
- cpu, 0, 0, 0, 0);
- continue;
- }
- callp->countp = NULL;
- } else {
- if (callp == NULL) {
- mp_call_head_unlock(cqp, intrs_enabled);
- KERNEL_DEBUG_CONSTANT(
- TRACE_MP_CPUS_CALL_NOBUF,
- cpu, 0, 0, 0, 0);
- if (!intrs_enabled) {
- /* Sniffing w/o locking */
- if (!queue_empty(&cqp->queue))
- mp_cpus_call_action();
- handle_pending_TLB_flushes();
- }
- cpu_pause();
- goto queue_call;
+ intrs_inner = mp_call_head_lock(cqp);
+ if (callp == NULL) {
+ mp_call_head_unlock(cqp, intrs_inner);
+ KERNEL_DEBUG_CONSTANT(
+ TRACE_MP_CPUS_CALL_NOBUF,
+ cpu, 0, 0, 0, 0);
+ if (!intrs_inner) {
+ /* Sniffing w/o locking */
+ if (!queue_empty(&cqp->queue))
+ mp_cpus_call_action();
+ handle_pending_TLB_flushes();
}
- callp->countp = &mp_cpus_calls;
+ if (mp_spin_timeout(tsc_spin_start))
+ panic("mp_cpus_call1() timeout start: 0x%llx, cur: 0x%llx",
+ tsc_spin_start, rdtsc64());
+ goto queue_call;
}
+ callp->maskp = (mode == NOSYNC) ? NULL : &cpus_responded;
callp->func = action_func;
callp->arg0 = arg0;
callp->arg1 = arg1;
mp_call_enqueue_locked(cqp, callp);
- mp_cpus_signals++;
+ cpus_call_count++;
cpus_called |= cpu_to_cpumask(cpu);
i386_signal_cpu(cpu, MP_CALL, ASYNC);
- mp_call_head_unlock(cqp, intrs_enabled);
+ mp_call_head_unlock(cqp, intrs_inner);
if (mode == SYNC) {
- mp_cpus_call_wait(intrs_enabled, mp_cpus_signals, &mp_cpus_calls);
+ mp_cpus_call_wait(intrs_inner, cpus_called, &cpus_responded);
}
}
}
+ if (topo_lock) {
+ simple_unlock(&x86_topo_lock);
+ ml_set_interrupts_enabled(intrs_enabled);
+ }
/* Call locally if mode not SYNC */
if (mode != SYNC && call_self ) {
KERNEL_DEBUG_CONSTANT(
TRACE_MP_CPUS_CALL_LOCAL,
- VM_KERNEL_UNSLIDE(action_func), arg0, arg1, 0, 0);
+ VM_KERNEL_UNSLIDE(action_func), VM_KERNEL_UNSLIDE_OR_PERM(arg0), VM_KERNEL_UNSLIDE_OR_PERM(arg1), 0, 0);
if (action_func != NULL) {
ml_set_interrupts_enabled(FALSE);
action_func(arg0, arg1);
}
/* For ASYNC, now wait for all signaled cpus to complete their calls */
- if (mode == ASYNC) {
- mp_cpus_call_wait(intrs_enabled, mp_cpus_signals, &mp_cpus_calls);
- }
+ if (mode == ASYNC)
+ mp_cpus_call_wait(intrs_enabled, cpus_called, &cpus_responded);
+
+ /* Safe to allow pre-emption now */
+ mp_enable_preemption();
out:
- cpu = (cpu_t) mp_cpus_signals + (call_self ? 1 : 0);
+ if (call_self){
+ cpus_called |= cpu_to_cpumask(cpu);
+ cpus_call_count++;
+ }
if (cpus_calledp)
*cpus_calledp = cpus_called;
- if (cpus_notcalledp)
- *cpus_notcalledp = cpus_notcalled;
KERNEL_DEBUG_CONSTANT(
TRACE_MP_CPUS_CALL | DBG_FUNC_END,
- cpu, cpus_called, cpus_notcalled, 0, 0);
+ cpus_call_count, cpus_called, 0, 0, 0);
- return cpu;
+ return (cpu_t) cpus_call_count;
}
static void
-mp_broadcast_action(void)
+mp_broadcast_action(__unused void *null)
{
/* call action function */
if (mp_bc_action_func != NULL)
/*
* signal other processors, which will call mp_broadcast_action()
*/
- simple_lock(&x86_topo_lock);
- mp_bc_ncpus = i386_active_cpus(); /* total including this cpu */
- mp_bc_count = mp_bc_ncpus;
- i386_signal_cpus(MP_BROADCAST, ASYNC);
+ mp_bc_count = real_ncpus; /* assume max possible active */
+ mp_bc_ncpus = mp_cpus_call(CPUMASK_OTHERS, NOSYNC, *mp_broadcast_action, NULL) + 1;
+ atomic_decl(&mp_bc_count, real_ncpus - mp_bc_ncpus); /* subtract inactive */
/* call executor function on this cpu */
- mp_broadcast_action();
- simple_unlock(&x86_topo_lock);
+ mp_broadcast_action(NULL);
- /* block for all cpus to have run action_func */
+ /* block for other cpus to have run action_func */
if (mp_bc_ncpus > 1)
thread_block(THREAD_CONTINUE_NULL);
else
lck_mtx_unlock(&mp_bc_lock);
}
+void
+mp_cpus_kick(cpumask_t cpus)
+{
+ cpu_t cpu;
+ boolean_t intrs_enabled = FALSE;
+
+ intrs_enabled = ml_set_interrupts_enabled(FALSE);
+ mp_safe_spin_lock(&x86_topo_lock);
+
+ for (cpu = 0; cpu < (cpu_t) real_ncpus; cpu++) {
+ if ((cpu == (cpu_t) cpu_number())
+ || ((cpu_to_cpumask(cpu) & cpus) == 0)
+ || (!cpu_datap(cpu)->cpu_running))
+ {
+ continue;
+ }
+
+ lapic_send_ipi(cpu, LAPIC_VECTOR(KICK));
+ }
+
+ simple_unlock(&x86_topo_lock);
+ ml_set_interrupts_enabled(intrs_enabled);
+}
+
void
i386_activate_cpu(void)
{
flush_tlb_raw();
}
-extern void etimer_timer_expire(void *arg);
-
void
i386_deactivate_cpu(void)
{
* and poke it in case there's a sooner deadline for it to schedule.
*/
timer_queue_shutdown(&cdp->rtclock_timer.queue);
- mp_cpus_call(cpu_to_cpumask(master_cpu), ASYNC, etimer_timer_expire, NULL);
+ mp_cpus_call(cpu_to_cpumask(master_cpu), ASYNC, timer_queue_expire_local, NULL);
/*
* Open an interrupt window
#if MACH_KDP
volatile boolean_t mp_kdp_trap = FALSE;
+volatile boolean_t mp_kdp_is_NMI = FALSE;
volatile unsigned long mp_kdp_ncpus;
boolean_t mp_kdp_state;
DBG("mp_kdp_enter()\n");
+#if DEBUG
+ if (!smp_initialized)
+ simple_lock_init(&mp_kdp_lock, 0);
+#endif
+
/*
* Here to enter the debugger.
* In case of races, only one cpu is allowed to enter kdp after
* "unsafe-to-interrupt" points such as the trampolines,
* but neither do we want to lose state by waiting too long.
*/
- tsc_timeout = rdtsc64() + (ncpus * 1000 * 1000 * 10ULL);
-
- if (virtualized)
- tsc_timeout = ~0ULL;
+ tsc_timeout = rdtsc64() + (LockTimeOutTSC);
while (mp_kdp_ncpus != ncpus && rdtsc64() < tsc_timeout) {
/*
* interrupt them with an NMI via the local APIC.
*/
if (mp_kdp_ncpus != ncpus) {
+ DBG("mp_kdp_enter() timed-out on cpu %d, NMI-ing\n", my_cpu);
for (cpu = 0; cpu < real_ncpus; cpu++) {
if (cpu == my_cpu || !cpu_datap(cpu)->cpu_running)
continue;
if (cpu_signal_pending(cpu, MP_KDP))
cpu_NMI_interrupt(cpu);
}
+ /* Wait again for the same timeout */
+ tsc_timeout = rdtsc64() + (LockTimeOutTSC);
+ while (mp_kdp_ncpus != ncpus && rdtsc64() < tsc_timeout) {
+ handle_pending_TLB_flushes();
+ cpu_pause();
+ }
+ if (mp_kdp_ncpus != ncpus) {
+ panic("mp_kdp_enter() timed-out waiting after NMI");
+ }
}
}
else
mp_kdp_wait(boolean_t flush, boolean_t isNMI)
{
DBG("mp_kdp_wait()\n");
- /* If an I/O port has been specified as a debugging aid, issue a read */
- panic_io_port_read();
+
current_cpu_datap()->debugger_ipi_time = mach_absolute_time();
#if CONFIG_MCA
/* If we've trapped due to a machine-check, save MCA registers */
debugger_exit_time = mach_absolute_time();
mp_kdp_trap = FALSE;
- __asm__ volatile("mfence");
+ mfence();
/* Wait other processors to stop spinning. XXX needs timeout */
DBG("mp_kdp_exit() waiting for processors to resume\n");
(void) ml_set_interrupts_enabled(mp_kdp_state);
postcode(0);
}
+
+#define TRAP_DEBUGGER __asm__ volatile("int3")
+
+kern_return_t
+DebuggerWithCallback(kern_return_t (*callback) (void*),
+ void *callback_context,
+ boolean_t proceed_on_sync_failure)
+{
+ simple_lock(&debugger_callback_lock);
+
+ struct debugger_callback callback_buf = {
+ .callback = callback,
+ .callback_context = callback_context,
+ .proceed_on_sync_failure = proceed_on_sync_failure,
+ .error = KERN_FAILURE
+ };
+
+ assert(debugger_callback == NULL);
+ debugger_callback = &callback_buf;
+
+ TRAP_DEBUGGER;
+
+ debugger_callback = NULL;
+
+ simple_unlock(&debugger_callback_lock);
+
+ return callback_buf.error;
+}
+
#endif /* MACH_KDP */
boolean_t
-mp_recent_debugger_activity() {
+mp_recent_debugger_activity(void) {
uint64_t abstime = mach_absolute_time();
return (((abstime - debugger_entry_time) < LastDebuggerEntryAllowance) ||
((abstime - debugger_exit_time) < LastDebuggerEntryAllowance));
* Cold start
*/
clock_init();
- cpu_machine_init(); /* Interrupts enabled hereafter */
- mp_cpus_call_cpu_init();
- } else {
- cpu_machine_init(); /* Interrupts enabled hereafter */
}
+ cpu_machine_init(); /* Interrupts enabled hereafter */
}
#undef cpu_number
{
cpu_warm_data_t cwdp = (cpu_warm_data_t)arg;
- timer_call_enter(cwdp->cwd_call, cwdp->cwd_deadline, TIMER_CALL_CRITICAL | TIMER_CALL_LOCAL);
+ timer_call_enter(cwdp->cwd_call, cwdp->cwd_deadline, TIMER_CALL_SYS_CRITICAL | TIMER_CALL_LOCAL);
cwdp->cwd_result = 0;
return;
return cwd.cwd_result;
}
}
+
+#if DEBUG || DEVELOPMENT
+void
+kernel_spin(uint64_t spin_ns)
+{
+ boolean_t istate;
+ uint64_t spin_abs;
+ uint64_t deadline;
+
+ kprintf("kernel_spin(%llu) spinning uninterruptibly\n", spin_ns);
+ istate = ml_set_interrupts_enabled(FALSE);
+ nanoseconds_to_absolutetime(spin_ns, &spin_abs);
+ deadline = mach_absolute_time() + spin_ns;
+ while (mach_absolute_time() < deadline)
+ cpu_pause();
+ ml_set_interrupts_enabled(istate);
+ kprintf("kernel_spin() continuing\n");
+}
+#endif
projects / apple / xnu.git / blobdiff